Signal extraction

The step that turns a raw customer interaction — a transcript, a ticket, a sales call — into discrete signals. Done well, it dedupes and filters; done poorly, it duplicates noise.

Signal extraction is the upstream step of synthesis: given a customer interaction (a meeting transcript, a support ticket, a sales call), pull out each individual signal — feature request, pain point, dismissed concern, aspirational comment, blocker. Each signal becomes a record the rest of the pipeline can cluster and count.

A single-pass extractor — read the transcript once, list the signals — produces noisy output. The same customer often raises the same issue three times in one call ("the export is slow," "exports really are slow," "if exports were faster…"); a literal pass yields three signals when there's one. Worse, off-hand curiosity ("could it ever do X?") and explicitly dismissed concerns ("I thought it didn't support Y, but I was wrong") get extracted as if they were real asks.

Kiln uses a two-pass approach. Pass 1 is liberal — it errs on the side of completeness and pulls every candidate signal, accepting that some will be noise. Pass 2 reviews each candidate in the context of the whole transcript: it dedupes signals that would be resolved by the same fix, suppresses dismissed concerns, and trims aspirational comments that aren't real product implications. The result is fewer, cleaner signals — typically half as many as a naive extractor, and worth several times more in clustering quality.